Research On Monocular Visual-inertial Localization System For Autonomous Driving In Closed Park

In recent years,the development of artificial intelligence has contributed to the fire of autonomous driving technology.The active participation of universities,research institutes and enterprises has enabled autonomous driving technology to gradually move from research on paper to on-the-ground applications.Enclosed parks have become a hot spot for current research due to low vehicle speed,simple road condition and few dynamic objects,and in order to achieve safe and stable driving operations,self-driving vehicles are required to have the ability to sense environment,position accurately and robustly.The current common positioning method for autonomous driving is Integrated Inertial Navigation System consisting of Global Navigation Satellite System(GNSS)and Inertial Navigation System(INS).With dual antenna positioning method,vehicles can achieve centimeter-level positioning if the signal is good.However,due to the influence of bad weather and tree shading,the signal produces jumps,and the absolute position error obtained at this point is large.Monocular visual-inertial localization systems can still provide reliable self-motion estimation in some challenging environments,which provides a new solution to the above problem.However,in the area of autonomous driving,the inaccuracy of solving inertial states during the initialization period of the monocular visual-inertial localization system,the deterioration of state estimation errors during the degenerate motions of the vehicle and the instability of the outliers have seriously affected positioning accuracy.In this paper,the monocular visualinertial localization system for autonomous driving in closed park is used to estimate all inertial states accurately at the same time by using initialization algorithm based on the maximum posterior,reduce the cumulative errors of the poses by using the multi-mode state estimation algorithm,and improve the outlier removal algorithm based on the normalized reprojection errors to achieve stable and efficient outlier removal in order to improve the vehicle localization under different driving conditions in the park.The major research components of this paper are formulated as follows.(1)Design of adaptive initialization algorithm based on maximum posteriori.In order to provide accurate inertial states to the back end of the monocular visual-inertial localization system,this paper derives the initialization problem of the monocular visual-inertial localization system as a nonlinear optimization problem based on the maximum posteriori.Firstly,IMU data with sufficient motion excitation and images satisfying parallax requirements are screened,and the initial estimation of inertial states is obtained from the image information.Then using the pre-integration formula to construct the initialization cost function and adding a priori for the accelerometer deviation to estimate all inertial states at once by Bundle Adjustment,reducing the errors brought by previous methods using decoupled estimation.(2)Design of multi-mode state estimation algorithm.To address the problem of cumulative errors in state estimation due to degenerate motions of vehicle,this paper classifies the vehicle motion and establishes three modes according to the observability of monocular visual-inertial localization systems in different situations.By constructing the mode judgment cost function,the errors of scale and gyroscope deviation are calculated to determine the specific mode to enter.Unobservable states under different modes are kept constant at this point to reduce the effect of errors in state estimation.Additionally,this paper uses PnP or IMU integration depending on the chosen mode to accurately find the states of the new frame.(3)Improved outlier removal algorithm based on normalized reprojection errors.When estimating relative poses between frames,translation and depth are more sensitive to error sources and prone to produce errors.This paper improves the outlier removal algorithm based on normalized reprojection errors,taking vehicle’s translation and feature point depth as states,constructing the reprojection errors.Then by decoupling the rotation to obtain the optical flow of each feature point caused by translation only,and finally using their normalized reprojection error as the evaluation criterion to obtain the inner point,which can effectively and stably improve the outlier removal effectiveness of the system.(4)The algorithm is verified by using real vehicle test in a closed park.This paper builds an autonomous driving platform vehicle,installing and debugging sensors,selecting a suitable park to collect multiple sets of data for trials and analysis.Compared with the prevailing monocular visual-inertial localization system VINS-Mono,the system of this paper has improved the accuracy of both rotation and translation,which fully proves the effectiveness of the paper’s algorithm.